“The use of parallel compute in models like O1 Pro is primarily for achieving robustness and predictable performance, not for searching low-probability tokens.”